canvascritic

Most healthcare systems are not using Azure, AWS, or GCP

This may be true for some large players in coastal states but definitely not true in general

Your typical non-coastal state run health system does not have model access outside of people using their own unsanctioned/personal ChatGPT/Claude accounts. In particular even if you have model access, you won't automatically have API access. Maybe you have a request for an API key in security review or in the queue of some committee that will get to it in 6 months. This is the reality for my local health system. Local models have been a massive boon in the way of enabling this kind of powerful automation at a fraction of the cost without having to endure the usual process needed to send data over the wire to a third party

Healthcare organizations that can't (easily) send data over the wire while remaining in compliance

Organizations operating in high stakes environments

Organizations with restrictive IT policies

To name just a few -- well, the first two are special cases of the last one

RE your hallucination concerns: the issue is overly broad ambitions. Local LLMs are not general purpose -- if what you want is local ChatGPT, you will have a bad time. You should have a highly focused use case, like "classify this free text as A or B" or "clean this up to conform to this standard": this is the sweet spot for a local model

SnipVex clipjacking wallets is almost beside the point, the real failure is a printer vendor treating software like a side gig. Printer and hardware companies get a pass on basic infosec hygiene that would be unacceptable for open source maintainers.

until that changes, airgap your weird hardware setups I guess

Also this is a perfect storm for lateral movement. USB-borne worms still work frighteningly well in small biz environments, especially ones with no centralized IT and people plugging printers directly into Windows desktops with admin perms. Here SnipVex is just a cherry on top-a nice, opportunistic payload for the growing class of infostealers targeting crypto wallets

Real question is whether this is just symbolic or if the French state will actually redirect procurement pipelines + vendor mandates around these principles. i'd be more impressed if this came bundled with policy teeth, e.g. requiring all software vendors to deliver open-by-default interfaces or pushing funding toward open infra maintenance. Otherwise it's hardly much more than a manifesto

I'm not a fan of neurophysiology analogies because it veer into pseudoscience, but I'll play along.

Roam implemented static bidirectional links and called it associative memory. in reality, it's closer to mind-mapping software with backlinks. So without mechanisms for reinforcement (surfacing old notes intelligently), pruning (forgetting irrelevant junk), or plasticity (reorganizing in response to use), the system becomes a junkyard of half-formed thoughts.

Brains forget for a reason, roam doesn't

this all makes sense and is honestly the most coherent humanoid startup thesis i've seen outside of figure.ai. You're right that the unit economics of hardware are a trap unless you can commoditize the complements. And humanoid hardware clearly wants to become a commodity, but no one's finished the job yet and it seems brutally difficult (see: the ghost of Willow Garage)

The tesla analogy makes sense to me but with a caveat: they still spend billions on CapEx and own verticals like battery chemistry and drivetrain design. In this case you’re betting that the value collapses upward into software, like the shift from phones to apps, but for that to work, your software has to deliver exponential delta per dollar

With that I think the real risk is that your "clean your house" package is deceptively hard in the long tail, and you will end up with the iRobot Roomba UX. Novelty fades fast when it constantly gets stuck under the couch or whatever the equivalent of that is for humanoids. To be fair iRobot/Roomba is a household name but still "only" a ~$1.5B company, which seems meager compared to ambitions in this space

As an aside I would love to see an RFC-style doc on how you think humanoid software standards should emerge. ROS is still a frankenstein, and someone needs to kill it gently lol

When my grandfather died, the only thing he asked to be buried with was a small pouch of roasted chestnuts. he used to say they reminded him of long cold walks home through wartime forests that smelled like smoke and bark.

Anyway after the funeral, I cracked one open by the fire and it was still sweet. RIP baba

Thanks for the reply

IME the COTS angle cuts both ways. It brings costs down and makes iteration faster, but whats the moat then?

if the value is in integration, that’s fine, but integration is fairly fragile IP. Open source is good reputationally but accelerates the diffusion of your edge unless the play is towards community+ecosystem lock-in or being the canonical reference impl (cf. ROS, HuggingFace)?

This is a clever hack and a cute abuse of SQL joins to brute-force what’s essentially a 2-ply MDP over a finite space.

The core idea btw of using precomputed transition/score tables to simulate and optimize turn-by-turn play is a classical reinforcement learning method

What would be interesting here is to flip it: train a policy network (maybe tiny, 2-layer MLP) to approximate the SQL policy. then you could distill the SQL brute-force policy into something fast and differentiable.

i’d love to see a variant where the optimizer isn’t just maximizing EV, but is tuned to human psychology. e.g., people like getting Yahtzees more than getting 23 in chance. could add a utility function over scores.

Anyway this is a great repo for students to learn expected value optimization with simple mechanics.

Not surprised to see this. Whats interesting to me in all this is the misplaced faith in emergent structure.

Roam bet on the idea that if you link enough atomic notes, structure will self-organize.

Which is such a weird fantasy if you spend a few minutes thinking about it. Try writing code like that or building a company or just about anything else! Why should notetaking and archive development be any different

It's clear you need some sort of editorial hand to create something maintainable and future proof. Like zettelkasten had Luhmann’s obsessive discipline behind it. Evidently roam had um. enthusiasm and javascript?

and yeah, it’s telling that the comparison is to IDEs. Imagine an IDE that dumped every snippet you typed into a graph database and expected you to recompile coherence out of it by browsing links. thats what roam felt like after the honeymoon.

In general most of Roam's target should want to lean harder into opinionated workflows. there’s a reason tools like linear or notion are winning. they’re structured enough to relieve cognitive load, flexible enough to adapt. Roam tried to be emacs, but turns out most users don’t want to configure their own productivity dialect.

also, lol at the idea of "automated taxonomy". The entire knowledge management industry keeps rediscovering ontologies like they’re new. We are probably going to reinvent OWL at some point and give it a name like "neuroschema" or something

Heh. This landing page takes me to somewhere between deepmind circa 2014 and tesla's AI Day press decks.

I mean if you're actually training humanoids in under an hour with sim-to-real transfer that "just works" then congrats, you've solved half of embodied AI

the vertical integration schtick (from "metal to model") echoes early apple, but in the robotics space that usually means either 1) your burn rate is brutal and you're ngmi, or 2) you're hiding how much is really off-the-shelf

Clearly the real play here, assuming it's legit, is the RL infra. K-Sim is def interesting if it's not just another wrapper over Brax/Isaac. Until we see actual benchmarks re say, dexterous manipulation tasks trained zero-shot on physical hardware, it's hard to separate "open-source humanoid stack" from the next pitch that ends in "-scale"

Clearly a labor of love. Props to you

I suppose if one is teaching or evangelizing constructor theory, this could be sort of like an interactive textbook

Needless to say, constructor theory hasn't really earned a stable foothold in mainstream physics, and there's a lot of hype in this space, but that's not a criticism of this particular project, just good to know for anyone not familiar

The quantum gravity + graviton tasks stuff especially. without a falsifiable physical model backing it, this can feel like mathematized cosplay. But that has more to do with constructor theory vs this project

Would love to see someone do a pluggable backend so you could test different "task ontologies" against each other.

Mainly I came here to say that categories can likely be used to great effect here a la Geroch

For instance you can start by modeling tasks as morphisms between substrate states (objects), and then enforce composition explicitly. define constructors as functors that map tasks and substrates while preserving structure.

for quantum or irreversible effects, use monads to encapsulate branching and decoherence. Then one could represent task sequences as categorical diagrams and check for commutativity. Or embed substrates via Yoneda to expose behavior in terms of available tasks

Right: Maya is a "language continuum" in the sense that geographically proximate speakers tend to understand each other well, and intelligibility goes down as you move further away from any given individual on the continuum

I use nearly of these tools with the exception of Perl. I go to great lengths to make sure I have access to them because it's so critical for quality of life. I love them and I understand why people love them.

Here's the reason: these languages/tools are tactically very powerful. Tactics are immediate and decisive. Tactics are effectively tricks in the sense that if you can "spot the trick", you can -- with a tiny amount of work -- reduce a formidable problem to virtually nothing. Having a vast toolkit that facilitates such tricks is incredibly powerful and makes you appear to have superpowers to colleagues who aren't familiar with them.

But tactics are definitionally short-term. You deploy them in the weeds, or at least from the forest, (hopefully) never from the skies. Tactics aren't concerned with the long term, nor how things fit together structurally. They are not concerned with maintainability or architecture.

This is why it isn't actually that important that you can cobble together a 15 line Perl script in an hour to do something that would take any of your colleagues a week. Years from now, when you are gone and someone runs into a similar but slightly different problem, someone will find your Perl script, not understand it, and rewrite it all in Java anyway. Or assume it's too hard and give up. Maybe they will adapt your Perl script, but more likely it'll be seen as a curiosity

It sucks, because there is beauty in that approach of solving problems. As I said in another comment, I wish there were more diversity in tooling and languages. But at the same time, it's important to consider that people are fundamental. All of this is in service to that. And I personally would rather build software that people use over the long term.